Robust Tensor Decomposition for Incomplete and Corrupted Data

This direction is merged into another direction: tensor decomposition scalability. Its last version stays readable and citable here.

Version 1 of 113 papers15 conceptsfrozen 2026-10-04

In many real-world applications—from traffic monitoring to multi-view video processing—data is naturally multidimensional. While we often flatten this data into matrices to use classical tools like Principal Component Analysis (PCA), this "unfolding" destroys the inherent structural correlations across different dimensions.

Recent research has converged on a more powerful approach: Robust Tensor Decomposition. This framework allows us to recover a clean, low-rank signal from observations that are simultaneously riddled with missing values and gross corruptions.

The Core Intuition: Low-Rank plus Sparse

The fundamental premise, as explored by Jonathan Q. Jiang and Michael K. Ng 0001 in Exact Tensor Completion from Sparsely Corrupted Observations via Convex Optimization, is that a corrupted observation $\mathcal{X}$ can be decomposed into two distinct parts:

  1. A Low-Rank Component ($\mathcal{L}$): This represents the underlying "ground truth" or signal. In a traffic sensor network, this is the consistent daily pattern of flow.
  2. A Sparse Component ($\mathcal{S}$): This represents outliers, sensor failures, or anomalies. These are "gross" errors that are large in magnitude but affect only a small fraction of the data.

By solving a convex optimization problem that minimizes the Tubal Nuclear Norm (TNN) of $\mathcal{L}$ and the $\ell_1$-norm of $\mathcal{S}$, researchers can mathematically guarantee the perfect recovery of the signal even when the data is highly incomplete.

Why the "Tubal" Perspective Works

Traditional tensor methods like CP or Tucker decomposition often struggle with computational complexity or lack clear convex relaxations. The breakthrough in this field comes from the Tensor Singular Value Decomposition (t-SVD) framework.

Instead of treating a 3D tensor as a block of independent numbers, the t-SVD treats it as a matrix of "tubes" (vectors extending through the third dimension). As Michael Ng and Qiang Jiang demonstrate in Robust Low-Tubal-Rank Tensor Completion via Convex Optimization, this algebraic structure allows us to apply matrix-like operations in the Fourier domain.

This perspective is powerful because it captures "circular convolution" effects along the third dimension, which is ideal for data with periodic or temporal dependencies. To make this even more flexible, Guang-Jing Song and colleagues in Robust Tensor Completion Using Transformed Tensor SVD showed that we can replace the Discrete Fourier Transform with any unitary transform, allowing the model to adapt to the specific "compactness" of different data types like hyperspectral images or face databases.

Capturing Local and Global Structure

While the low-rank constraint captures global patterns, real-world data often has local smoothness that simple rank-minimization misses.

In the context of Intelligent Transportation Systems, Jicheng Li introduced the RTC-GTNLN model, which uses a "Gradient Tensor Nuclear Norm." This ensures that the recovered traffic data is not just globally consistent (low rank) but also locally smooth (small gradients), preventing jittery reconstructions.

Similarly, for time-series anomalies, Xudong Wang and Lijun Sun 0001 proposed Hankel-structured Tensor Robust PCA. By transforming a 1D or 2D time series into a higher-order "Hankel tensor," the model can exploit the low-rank nature of linear dynamical systems, making it much easier to spot anomalies that deviate from those dynamics.

Application: Multi-View Clustering

One of the most successful applications of this idea is in Incomplete Multi-View Clustering. When we have multiple sources of information (e.g., audio, video, and text) but some sources are missing for certain samples, we can stack the similarity graphs from each view into a 3D tensor.

As Chao Zhang explains in Enhanced Tensor Low-Rank and Sparse Representation Recovery, enforcing a low-rank constraint on this "similarity tensor" allows the model to "borrow" information from available views to fill in the gaps of missing ones. Zhengqiao Zhao further refined this in JTIV-LRR by enforcing low-rank constraints across all three dimensions simultaneously, ensuring that the relationships between different views are just as consistent as the relationships within a single view.

Overcoming Orientation Bias

A final piece of the puzzle is orientation invariance. Standard tubal-rank methods are sensitive to which dimension is chosen as the "third" dimension. Andong Wang and Qibin Zhao addressed this in Robust Tensor Decomposition via Orientation Invariant Tubal Nuclear Norms. By utilizing a "3d-unfolding" technique, they created norms that exploit low-rank structures across all possible orientations at once. This ensures that whether a correlation exists horizontally, vertically, or temporally, the robust decomposition framework will find it and use it to purge the noise.

Papers in this version

History & citationVersion 1 · 2026-10-04

Merged into another direction · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_robust_tensor_decomposition_for_incomple_df522eea,
  title = {Robust Tensor Decomposition for Incomplete and Corrupted Data},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/robust-tensor-decomposition-for-incomplete-and-corrupted-data/rdr_df522eea897943859b1d7e1ef7afac00},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-19},
}

This exact version (version 1, frozen text):

@misc{lacuna_robust_tensor_decomposition_for_incomple_df522eea_v1,
  title = {Robust Tensor Decomposition for Incomplete and Corrupted Data},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/robust-tensor-decomposition-for-incomplete-and-corrupted-data/rdr_df522eea897943859b1d7e1ef7afac00/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-19},
}